We present a new approach to enhance the performance of trading strategies developed through deep reinforcement learning algorithms in the highly unpredictable environment of intraday cryptocurrency portfolio trading. Our method involves using an ensemble technique to improve the generalizability of these strategies. To achieve this, we employ a model selection process that assesses performance across multiple validation periods. Additionally, we introduce a unique mixture distribution policy that effectively combines the selected models for optimal results.
To demonstrate the robustness of our strategies in varying market conditions, we analyze the out-of-sample performance on granular test periods from a distributional perspective. We also address the non-stationarity of financial data by periodically retraining the models.
Through our proposed ensemble method, we are able to enhance the performance of the strategies in comparison to the benchmarks set by a deep reinforcement learning strategy and a passive investment strategy.